Continual Test-Time Adaptation

Continual test-time adaptation studies how pretrained models can adjust to unlabeled data streams as conditions change, without relying on repeated access to training data or labels. Common approaches use entropy-based objectives, pseudo-labels, prototypes, memory and replay, or lightweight prompts and adapters; newer methods also detect shifts and adapt selectively or without backpropagation. The field spans vision and segmentation as well as forecasting, graphs, multimodal systems, and deployed agents, where adaptation can improve performance under domain shift. Its central challenge is balancing responsiveness with stability: noisy updates can cause error accumulation and forgetting, so long-horizon evaluation, resource costs, and comparison with frozen models remain important.